I mean, if the test user can't figure it out at all, how is the rest of the UI/documentation supposed to get evaluated?

Great question!

If you let someone flounder on one task indefinitely then you don't learn anything about subsequent tasks. But if you correct them too quickly you won't uncover the other approaches they would have tried to complete the task. Most research plans define cutoffs such as:

1. Participant expresses extreme frustration or gives up

2. A couple minutes have elapsed from the first failed attempt

3. Participant unsuccessfully attempts three distinct approaches

If the test reaches one of your cutoffs then the interface/docs have failed the task and the moderator can skip to the next task or question. Sometimes they'll also offer to show the participant the expected solution or explanation.

Exactly. You want to learn as much as possible from each study. Explaining too soon reduces amount learned, as does ending the study early because a small hint wasn't provided to get to the next step.